This week in AI — Aug 17 – 23, 2026

DeepSeek has launched vision capabilities on its V4 flash model, which is now live and available for use. Users are already running DeepSeek V4 flash on platforms like DeepInfra, noting it as a more economical option compared to the official API, especially without peak-time pricing. Further details can be found in the DeepSeek official API documentation.

GPT-5.6 Sol, the frontier model in the GPT-5.6 family, now offers a 20% reduction in input pricing and a 33% reduction in output pricing, costing $4 per million input tokens and $20 per million output tokens. This model is designed for complex professional work, featuring high reasoning capabilities, fast speed, and a 1,050,000 context window. It supports text and image input, text output, and has a knowledge cutoff of February 16, 2026. Promotional pricing is available at least through November 21, 2026.

A new tool, tentatively named "Claudette," has been developed to transform Claude's responses from a "millennial clickbait" style to regular English. This skill, accessible via /debuzz, processes Claude's last response through the Antigravity CLI (agy). The developers humorously suggest that Claude's conversational style is due to being trained on old BuzzFeed articles, explaining its "love for 90s nostalgia."

The Qwen 3.8 27B model, running on a Lenovo ThinkStation PGX with Nvidia's GB10 Grace Blackwell chip, demonstrated impressive performance in a reverse-engineering task, completing it in 30 minutes. While initially achieving 15 to 30 tokens per second, a speculative-decoding setup using SGLang, NVFP4, and DFlash2 boosted its speed to around 50 tokens per second. Artificial Analysis ranks Qwen 3.8 27B as the top open-weights model in its 4B to 40B size class, outperforming 135 other models with a 52 on its intelligence index.

Vomit is a fully local, open-source tool (GNU GPLv3) designed to convert Claude's "token vomit" into English using a separate local LLM, such as Llama.app or Ollama. It operates without telemetry or external dependencies, offering modes like `vomit scrub -claude` to replace Claude's output or `vomit tail` for non-invasive translation. While it may hallucinate and is currently tested only on Mac, it provides a way to clarify Claude's communications.

This research paper, titled "Universality of Gradient Descent Neural Network Training," explores machine learning concepts. It was first made available on July 27, 2020, as version 1 (v1) and is identified by arXiv:2007.13664 [cs.LG]. The document is 31 KB in size and can be cited using its DOI, 10.48550/arXiv.2007.13664. The primary subjects are Machine Learning (cs.LG) and Machine Learning (stat.ML).

Wiz Research, using its AI tool "Red Agent," discovered a critical GitHub Actions workflow vulnerability in a Snowflake public repository. The vulnerability, introduced by commit 094038e and merged via PR #1218, involved an injectable pattern in jira_issue.yml. GitHub Copilot Autofix contributed a separate fix in the same PR, but GitHub Advanced Security failed to flag the injection. Snowflake remediated the issue reported on June 23, 2026, finding no evidence of unauthorized access, and is collaborating with Wiz to share these security learnings.

OpenAI recently launched the GPT-5.6 lineup, including the Sol, Terra, and Luna models, with a strong focus on computer vision capabilities. While GPT-5.6 Sol demonstrated a 90.7% mean similarity score in OCR, slightly behind GPT-5.5's 91.2%, and 82.5% in text extraction compared to GPT-5.5's 87.6%, it signifies OpenAI's increased commitment to vision. Despite some flaws in cost, latency, and detection, this release significantly enhances OpenAI's offerings for agents, screen understanding, and visual reasoning.

Speko (YC S26) is introduced as an "OpenRouter for Voice AI." It functions as a router for voice AI, utilizing components like `openai.STT`, `openai.LLM`, and `openai.TTS` within a `voice.AgentSession`. The provided code snippet shows an agent definition using `defineAgent` and a command-line interface example `claude mcp add --transport http speko https://mcp.speko.ai/mcp` for integration.

This week, the author used Codex more than Claude, noting that Codex produced simpler code architecture. Claude tended to create more abstractions and complex solutions, though its code handled more cases. Codex, in contrast, was more contained and acted as a companion, doing exactly what was asked without overdoing it. Claude, however, often tried to anticipate and implement additional user needs.

1667 is a terminal UI designed for writing fiction using language models. It allows users to create and manage their written content, ensuring that every piece of writing is preserved. The tool can be installed via a script downloaded from GitHub releases, specifically version v0.9.8, with attestation verification for security. Installation commands are provided for both Unix-like systems and PowerShell.

Claude Code is offering a limited-time promotion from May 13, 2026, through August 19, 2026, at 11:59 PM PT. This promotion increases weekly usage limits in Claude Code by 50%. It applies to Pro, Max, and Team plans, as well as legacy seat-based users on Enterprise plans. Free plans and consumption-based Enterprise seats are excluded. This offer has no cash value, is not transferable, and cannot be combined with other offers.

OneCLI (YC S26) is an open-source sandboxed agent harness designed for teams. It operates under an Apache-2.0 license, with the exception of enterprise features located in `ee/` directories. These enterprise features are governed by the OneCLI Enterprise License, which permits free use for development, testing, and evaluation, but requires a subscription for production use. All other components are Apache-2.0 licensed and can be self-hosted in production without a commercial license.

Anthropic's annualized revenue reached $65bn in July, up from $47bn in May, with Q3 expected to be profitable. They boast 6,000 customers spending over $100,000 annually. Meanwhile, OpenAI's annualized revenue surpassed $40bn, boosted by the July launch of GPT 5.6. Despite this growth, Anthropic's Fable 5 model struggles with only 8.0% of model spend, while Opus 4.8 leads with 28.0%, suggesting that Fable's cost may hinder its adoption.

OpenAI's Q2 revenue announcement coincides with a broader market downturn, as AI and chip stocks experience declines amidst US-Iran tensions. Meta Platforms is also involved in a lawsuit trial, contributing to the overall market volatility. The Basis Points Pod with Amit discusses these developments, noting that hosts and guests may have financial interests in the companies or assets mentioned, and past performance does not guarantee future results.

A developer trained a 125M-parameter transformer model to autocomplete piano performances in real time, achieving approximately 108 notes/sec on an iPhone 15. Key improvements stemmed from optimizing the MIDI representation, aggressive data cleaning, and implementing DPO post-training. Pairwise evaluation using Gemini 3.5 Flash helped build a preference dataset for DPO, by asking it to compare continuations rather than assign absolute scores, and mirroring comparisons to reduce position bias.

AI chip startup Etched has secured $700 million in a recent funding round, elevating its valuation to $21 billion. This significant investment highlights the growing interest in specialized AI inference chips. The development raises questions about whether these chips can challenge Nvidia's current dominance in the AI market, or if GPUs will continue to be the primary choice for most AI workloads, indicating a potential infrastructure battle.

AI capabilities are advancing rapidly, posing risks to humanity's survival, autonomy, and democratic order, especially if power and wealth concentrate with AI system owners. OpenAI, once a non-profit, is now a for-profit corporation after its non-profit commitment was scrapped. It is suggested that Norway, using its GPF-G fund valued at over $2 Trillion, should buy OpenAI, valued around $800 billion, to return the lab to public hands, despite requiring Norges Bank Investment Management to liquidate 40% of its portfolio and violate the GPF-G’s mandate.

OpenAI has halted training and evaluations for its forthcoming frontier AI model, Astra, to implement new cybersecurity risk procedures. This decision was prompted by an internal evaluation showing Astra's enhanced coding and cybersecurity abilities, the general pace of AI progress, and incidents like the Hugging Face saga, which revealed the company had underestimated its AI models' real-world cyber capabilities. OpenAI is introducing new monitoring, security, and alignment requirements to address these advanced hacking abilities.

OpenAI is reportedly facing internal turmoil, with a significant $7 billion insider cash-out raising concerns. The company has seemingly avoided a valuation test, and an executive exodus is worsening. Warning signs around Sam Altman are emerging, especially as individuals who challenged him are no longer with the company. Furthermore, OpenAI's safety guardrails are reportedly disappearing, leading to speculation that insiders might be selling at the top, despite a $400 billion bet riding on the company.

The GLM-5.3 open-weight model has reportedly outperformed models from Anthropic and OpenAI, achieving a 100% pass rate and a 9.3 rubric score at approximately one-fifth the cost of gpt-5.5. While gpt-5.5 offers faster Time To First Token (TTFT) at 13.2s compared to GLM-5.3's 16.3s, GLM-5.3 is presented as a cost-effective option. Other models like gpt-5.6-luna are noted for low-risk tasks, haiku-4-5 for accuracy, and sonnet-4-6 for quality and speed.

OpenAI is offering Zero Data Retention for eligible API customers, ensuring that prompts and model responses are not retained after processing. Customer content is not accessible to OpenAI personnel for review, and enterprise customer data is not used for model training unless customers explicitly opt-in. OpenAI plans to roll out Private Safety Processing and release a technical white paper in September, keeping customers informed about these updates and their implications for existing commitments.

OpenAI has introduced "ChatGPT for Teens," a version of its AI designed for users aged 13-17, featuring enhanced safety protections and controls for parents. This initiative aims to help teens learn, think critically, and use AI confidently, while promoting healthy usage. Examples of AI's positive impact include WiFind, a search-and-rescue system developed by teens using Wi-Fi signals, and Audemy, an educational game platform for blind students scaled to 200,000 users with ChatGPT's assistance. OpenAI emphasizes providing access to AI with protections tailored to this developmental stage.

OpenAI has introduced a version of ChatGPT tailored for teenagers, incorporating new safeguards to ensure user safety. Chris Lehane, OpenAI's Chief Global Affairs Officer, discussed these measures with Stephanie Ruhle. This development aims to provide a secure AI experience for younger users, addressing concerns about content and interaction. The announcement highlights OpenAI's commitment to responsible AI deployment, particularly for a vulnerable demographic.

A man suspected a court of using AI and attempted to win his case by injecting prompts into his filings. This appears to be the first US instance of prompt injection in the court system. A similar case in Brazil involved two attorneys who used the same attack in an AI-reviewing court, resulting in monetary sanctions of about $16,000.